Skip to content

Latest commit

 

History

History
182 lines (134 loc) · 7.86 KB

README.md

File metadata and controls

182 lines (134 loc) · 7.86 KB

Rails Simpler Bench

Build Status

Are you looking for a comprehensive, real-world benchmark to show the performance of a large, highly-concurrent Ruby on Rails application? You're probably looking for Rails Ruby Bench. But RRB is unfortunately large and unwieldy - it's a real-world benchmark with lots of dependencies and messiness. The easiest way to run it is to build an AWS image and run it in a dedicated virtual machine. While RRB does what it can to be repeatable, it has some strong limits due to concurrency - the behavior of thread interaction can be highly unpredictable.

As well, RRB is based on Discourse, a real-world production application. This makes it hard to change Rails (and sometimes Ruby) versions, and it requires the benchmark to be fairly messy -- it has to follow the evolution of a real-world solution to real-world (unrelated to the benchmark) problems.

RRB's strengths and weaknesses are easy to sum up in the same sentence: it embraces and measures as much real-world complexity as it reasonably can.

Would you like a much-simplified Ruby on Rails benchmark with fewer dependencies and more repeatability? Then RSB may be for you.

This benchmark, like Rails Ruby Bench, was written and maintained via sponsorship from AppFolio (https://engineering.appfolio.com). Thank you, AppFolio!

Usage

Quick Start

There are a small, specific number of Rubies that are already supported (have a Gemfile.lock.) You can see which ones by looking in the rails_test_app or rack_test_app directory - but they include 2.0.0p0, 2.0.0p648, 2.1.0, 2.2.10, 2.3.8, 2.4.5, 2.5.3 and 2.6.0. JRuby 9.2.0.0 is also partially supported - see below in this file for more details.

You can add support for another Ruby by adding a Gemfile for it - see the existing Gemfiles for examples, but they're quite simple.

(Note: there is an experimental branch which allows dynamically generating Gemfiles for any Ruby)

If you want to check the current speed of a benchmark on your current machine, you can do it fairly simply:

./runners/current_ruby.rb

If you want to change settings, you can do it in that runner script.

You can also run with more settings changes and multiple different Rubies if you have RVM installed - see runners/rvm_rubies.rb for a full list of environment variables to set parameters. Here is an example, specifying many of them:

RSB_NUM_RUNS=10 RSB_RUBIES="2.6.0 2.4.5 2.0.0-p0" RSB_DURATION=180 RSB_WARMUP=20 RSB_FRAMEWORKS=rack RSB_APP_SERVER=puma RSB_PROCESSES=1 RSB_THREADS=1 ./runners/rvm_rubies.rb

Process Structure

  • Runner (e.g., runners/current_ruby.rb)
    • ruby wrk_subprocess.rb benchlib.json
      • Server (e.g., bundle exec rails server)
      • wrk OPTIONS URL

Runners

RSB uses a number of "runner" scripts to test different configurations. For most uses of RSB, you'll run one. For more customized uses of RSB, you'll make your own runner script.

The file runners/rvm_rubies.rb is the most comprehensive as of this writing - see the beginning of the file for documentation. It runs primarily from environment variables.

Analysis

After the runner completes, you should have a directory of data files, which can be used directly or analyzed. The file process.rb in the root directory of this repository performs cohort-based analysis, good for checking simple A/B questions of the form, "did this change speed up Ruby or slow it down? By how much?"

For instance, to compare the speed of many different Rubies using a directory of data files, you can often type something like "../process.rb -c 'rvm current'", which will use the recorded RVM Ruby for each batch of data to analyze each Ruby's subset of the data separately.

Experiments

The analysis above is okay. You can keep data files in multiple directories, analyze them separately and compare. But that's not always a reasonable choice, and often you'd rather do it differently.

The RSB output files record any environment variable containing "RUBY", "RSB" or "GEM" as relevant. If you want to conduct an experiment between multiple configurations, it's often a good idea to set an environment variable starting with RSB so that you can separate the relevant data files afterward. For instance, if you set "RSB_MY_EXP_CONFIG=7" beforehand, then that setting would propagate to the data file since the environment variable contains "RSB".

To group them by the value of that variable afterward, set up cohorts using that variable. For instance:

cd my_data_dir
~/src/rsb/process.rb -c "RUBY_VERSION,env-RSB_MY_EXP_CONFIG"

The above example would group the data into cohorts according to which Ruby version was used and the value of the environment variable RSB_MY_EXP_CONFIG. You can then compare the latencies, throughputs or other relevant data.

If for some reason setting an environment variable is inappropriate to your use case, you'll need to separate the relevant data in some other way.

JRuby

JRuby is intended to be a target Ruby configuration for RSB, but not everything is working yet.

Here are some restrictions:

  • JRuby can be specified as a target configuration, such as in RSB_RUBIES in runners/rvm_rubies.rb. However, some of the test harness can't run in JRuby because it doesn't have "fork". By starting the test harness from a CRuby implementation, you can still check JRuby's speed.

  • JRuby uses its own SQLite3 adapter. This will give slightly different performance than the sqlite3 gem, which isn't supported in JRuby.

Canonical Configurations

RSB will cheerfully run with whatever concurrency you like. However, here are some useful configurations to use or test:

  • 1 Process, 1 Thread: this gives minimum latency
  • N Processes, 1 Thread: high throughput for most workloads; you usually want roughly 1 process per core, or 1.3-1.5 processes/core with hyperthreaded cores
  • 1 Process, N Threads: this is usually best for JVM-based Ruby implementations like JRuby or TruffleRuby

Threads and Concurrency

An interesting property of RSB, especially compared to RRB or other large Rails apps, is that it isn't very threading-friendly on CRuby. The GIL means that Ruby code can't run concurrently with other Ruby code. A since RSB doesn't use Reddit or caches at all, or the database much, that means there's relatively little non-Ruby code.

For instance, with the static route, here are some example route throughputs:

(Format is # of Processes, # of threads, throughput in iters/sec, StdDev)

  • 1, 1, 1005 StdDev 11.4
  • 1, 2, 873, StdDev 17.1
  • 1, 3, 898, StdDev 8.9
  • 1, 4, 893, StdDev 5.0
  • 4, 1, 3892, StdDev 26.4
  • 4, 2, 3009, StdDev 20.7
  • 4, 4, 3289, StdDev 28.5
  • 4, 6, 3247, StdDev 51.6
  • 4, 8, 2901, StdDev 111.2
  • 8, 1, 5059, StdDev 118.3
  • 8, 2, 4776, StdDev 32.2
  • 8, 4, 4936, StdDev 209.7
  • 8, 6, 5039, StdDev 105.4
  • 8, 8, 5159, StdDev 178.8

Notice that in many cases, increasing the number of threads causes the throughput to go downward? RSB isn't particularly thread-friendly, especially when not using the database in the measured route.

Load Testing Tools

ApacheBench is no longer being used for RSB, and any remaining vestiges of it are just that. It doesn't report individual timings below millisecond resolution, its KeepAlive code is HTTP 1.0-only and causes bugs in Puma, and it has trouble with dynamic documents (cases where responses aren't byte-for-byte) identical. After extensive attempts to use it, the accuracy issues have been insurmountable (http://engineering.appfolio.com/appfolio-engineering/2019/1/18/benchmarking-ruby-app-servers-badly).

For newer use cases, we follow Phusion.nl's recommendation for "wrk", a simple, powerful and accurate benchmarking program that avoids reopening connections (Google "ephemeral port exhaustion" for details on why reopening connections gets to be a problem quickly on Linux.)